ainotis Join
My notis

Models

Every model our stories named. Each number links to its story and source. A dot means not checked yet.

All makers AlibabaAnthropicCanvaCloudflareElevenLabsGoogleH2O.aiLiquid AIMetaMicrosoftMicrosoft AIMiniMaxMistralMoonshot AINvidiaOpenAIReflection AISpaceXSIZ.ai
ModelMakerLatestPer million tokens, in / outContext Maker's own benchmarksMeasured by others
Claude Anthropic 15 Nov not checked yet not checked yet not checked yet not checked yet
Claude Mythos Anthropic 29 Oct not checked yet not checked yet not checked yet not checked yet
Claude Haiku 5.5 Anthropic 15 Oct $0.50 / $2.50 1,000,000 tokens Anthropic political evenhandedness evaluation 99% · OSWorld 2.1, offline subset 72.4% not checked yet
Claude Opus 5.5 Anthropic 15 Oct $4 / $20 1,000,000 tokens Anthropic political evenhandedness evaluation 99% Terminal-Bench 4.0 (as reported by Moonshot) 66.4% · CheatBench cheating rate with Claude Code (lower is better) 11.2%
Claude Sonnet 5.5 Anthropic 15 Oct $2 / $10 not checked yet Anthropic political evenhandedness evaluation 99% · OSWorld 2.1, offline subset 83.9% · Terminal-Bench 4.0 70.6% not checked yet
Claude Haiku 4.5 Anthropic 8 Oct $1 / $5 not checked yet OSWorld 2.1, offline subset 15.7% not checked yet
claude-mythos-preview Anthropic 1 Oct not checked yet not checked yet Binary Exploitation (Anthropic internal, 100 tasks) 6% not checked yet
Claude Opus 5 Anthropic 28 Sep not checked yet not checked yet not checked yet ProgramDistill cumulative workflows 28.8%
Claude 3.7 Sonnet Anthropic not checked yet not checked yet SHADE-Arena overall sabotage success, Claude 3.6 Sonnet as monitor 27% not checked yet
Claude Fable 5.1 Anthropic $10 / $50 not checked yet not checked yet not checked yet
Claude Fable 5.2 Anthropic $8 / $40 not checked yet not checked yet not checked yet
Claude Mythos 5.2 Anthropic $8 / $40 not checked yet dedicated sandbox escape evaluation, credentials or files used outside the sandbox (Anthropic's own test) 3.1% not checked yet

Maker's own: what a company measured on its own model, from its model card, system card or announcement. A claim about its own product, and labelled so.

Measured by others: a safety institute, a research group or a public benchmark run by someone other than the maker.

Prices for prompts over 100,000 tokens where a model has two tiers. Benchmarks come from each release's own documents, collected by the registrar.

Models

Every model our stories named. Each number links to its story and source. A dot means not checked yet.

All makers AlibabaAnthropicCanvaCloudflareElevenLabsGoogleH2O.aiLiquid AIMetaMicrosoftMicrosoft AIMiniMaxMistralMoonshot AINvidiaOpenAIReflection AISpaceXSIZ.ai

ModelMakerLatest in our stories Price per million tokens, in / outContext windowIndependent test results
Claude Anthropic 15 Nov · · ·
Claude Mythos Anthropic 29 Oct · · ·
Claude Haiku 5.5 Anthropic 15 Oct $0.50 / $2.50 1,000,000 tokens ·
Claude Opus 5.5 Anthropic 15 Oct $4 / $20 1,000,000 tokens Terminal-Bench 4.0 (as reported by Moonshot) 66.4% · CheatBench cheating rate with Claude Code (lower is better) 11.2%
Claude Sonnet 5.5 Anthropic 15 Oct $2 / $10 · ·
Claude Haiku 4.5 Anthropic 8 Oct $1 / $5 · ·
claude-mythos-preview Anthropic 1 Oct · · ·
Claude Opus 5 Anthropic Outside test 28 Sep · · ProgramDistill cumulative workflows 28.8%
Claude 3.7 Sonnet Anthropic · · ·
Claude Fable 5.1 Anthropic $10 / $50 · ·
Claude Fable 5.2 Anthropic $8 / $40 · ·
Claude Mythos 5.2 Anthropic $8 / $40 · ·